A speech recognition apparatus includes a speech collating unit that calculates similarities at each time between a feature amount converted by a speech analyzing unit and a word model generated by a word model generating unit. The speech collating unit extracts a word model from word models generated by the word model generating unit, whose minimum similarity among similarities at each time or whose overall similarity obtained from similarities at each time satisfies a second threshold value condition, and whose similarity at each time in a section among vocalization sections of utterance speech and corresponding to either a phoneme or a phoneme string associated with a first threshold value condition satisfies the first threshold value condition, and outputs as a recognition result the recognized word corresponding to the extracted word model.